TaylorSwiftNet

This repository contains the code for the paper Taylor Swift: Taylor Driven Temporal Modeling for Swift Future Frame Prediction.

Installation

Setup a conda environment and install all project dependencies.

conda env create --name taylor --file environment.yml
activate taylor
pip install -e .

How to run the code

To train the MovingMNIST model, use

python core/main.py --cfg configs/moving_mnist/latest_config.yaml \
--set dataset.root <path_to_dataset>

All config parameters are described in configs/default_config.py. You can specify parameters by setting them in a yaml config file or by passing them after --set (Format: --set <key1> <value1> <key2> <value2> ...).

To evaluate a previously trained model checkpoint, use

python core/main.py --cfg configs/moving_mnist/latest_config.yaml \
--set dataset.root <path_to_dataset> eval_only True model.resume True model.model_state_path <path_to_checkpoint.pt>

Citation

If you use this code or our models, please cite our paper:

@inproceedings{taylor2022,
    Author    = {Saber Pourheydari, Emad Bahrami, Mohsen Fayyaz, Gianpiero Francesca, Mehdi Noroozi, Juergen Gall},
    Title     = {TaylorSwiftNet: Taylor Driven Temporal Modeling for Swift Future Frame Prediction},
    Booktitle = {British Machine Vision Conference (BMVC)},
    Year      = {2022}
}

Contributors

Saber Pourheydari

Emad Bahrami

Mohsen Fayyaz

Acknowledgment

Felix helped us for refactoring and cleaning the original code.

Felix B. Müller